GPT-5.6 Terra

GPT-5.6 Terra (API model ID gpt-5.6-terra) is OpenAI's balanced mid-tier model in the three-tier GPT-5.6 family — Sol (flagship) and Luna (fastest/cheapest) are the other two — reaching general availability on 2026-07-09 after a gated preview that OpenAI began restricting to a small set of partner companies at the U.S. government's request in June 2026, before lifting the restriction ahead of the public launch (CNBC; GitHub Changelog). OpenAI positions Terra as the successor to the "mini" tier from earlier GPT-5 generations: "competitive with GPT-5.5" at roughly half Sol's per-token price (OpenAI API model docs). The bare gpt-5.6 alias in the API resolves to Sol, not Terra — pin the full gpt-5.6-terra ID to avoid silently paying flagship rates (OpenAI API model docs — gpt-5.6-sol).

Key specs

Values as of the 2026-07-09 GA release, cross-checked 2026-07-12 against the OpenAI API model docs and the models.dev snapshot (data/models/2026-07-12-models-dev.json) — the two sources agree on every field below, no discrepancy to flag.

Spec	Value	Source
Context window	1,050,000 tokens	OpenAI API model docs
Max output tokens	128,000	OpenAI API model docs
Modalities	Text + image input; text output only (no audio/video output)	OpenAI API model docs
Knowledge cutoff	2026-02-16	OpenAI API model docs
Input price	$2.50 / Mtok	OpenAI API model docs
Cached input price	$0.25 / Mtok	OpenAI API model docs
Output price	$15.00 / Mtok	OpenAI API model docs
Fine-tuning	Not supported	OpenAI API model docs
Rate limits, Tier 5	15,000 RPM / 40,000,000 TPM / 15,000,000,000 batch queue	OpenAI API model docs
Preparedness rating	Cybersecurity: High; Biological/Chemical: High; AI Self-Improvement: below High (all below "Critical")	Deployment Safety Hub system card
Availability	Responses/Chat Completions API; GitHub Copilot (Pro and above) across VS Code, JetBrains, Copilot CLI, and more; not selectable in standard ChatGPT conversations on any plan — selectable instead inside ChatGPT Work (Plus/Pro/Business/Enterprise) and Codex (Terra on Free/Go; a Sol/Terra/Luna choice with a reasoning-level picker on Plus/Pro/Business/Enterprise)	GitHub Changelog; OpenAI Help Center

Tool support over the Responses API includes web search, file search, image generation, code interpreter, hosted shell, apply-patch, skills, computer use, MCP, and tool search, plus Programmatic Tool Calling — model-generated JavaScript run in an isolated V8 sandbox with no network access, available at the same capability level on all three tiers (OpenAI API model docs; MarkTechPost, 2026-07-09).

Benchmark results

Benchmark	Terra	Sol (context)	Source	As of
SWE-bench Pro, OpenAI's own eval table	63.4%	64.6%	OpenAI GPT-5.6 announcement	2026-07-12
SWE-bench Pro, CodeRabbit independent harness (100+ tasks, TS/Go/Python/JS/Rust)	40.7% pass rate	63.7% pass rate	CodeRabbit blog	2026-07-12
SWE-bench Verified, Vals AI leaderboard	75.2%	96.2%	Vals AI	2026-07-12
Terminal-Bench 2.1, according to OpenAI (no independent Terra run on the tbench.ai leaderboard as of this writing)	87.4%	88.8%	OpenAI GPT-5.6 announcement	2026-07-12
Agents' Last Exam, according to OpenAI	50.4%	52.7%	OpenAI GPT-5.6 announcement	2026-07-12
OSWorld 2.0, according to OpenAI	50.2%	62.6%	OpenAI GPT-5.6 announcement	2026-07-12
ARC-AGI-2, 120 tasks, max reasoning effort, official ARC Prize leaderboard	83.9%	92.5%	ARC Prize	2026-07-12
CodeRabbit code-review benchmark, production PRs	52.5% pass rate (53/101)	69.7% pass rate (69/99)	CodeRabbit blog	2026-07-12

Reading notes for practitioners:

• SWE-bench Pro is the headline number, and it's contested. Independent analyst Simon Willison notes OpenAI's own audit found roughly 30% of SWE-bench Pro tasks contain errors, which bounds how much weight any single percentage on this benchmark deserves (Willison, 2026-07-12). On the SWE-bench Verified leaderboard, a different benchmark from SWE-bench Pro, Terra trails Sol by more than 20 points (75.2% vs 96.2%) — the "mid-tier at half the price" positioning holds for price but not for headline coding accuracy on that leaderboard (Vals AI, 2026-07-12).
• ARC-AGI-2 shows a wide tier gap on abstract reasoning. Terra scores 83.9% against Sol's 92.5% at the same max reasoning effort — an 8.6-point spread, though not the widest in this table: OpenAI's own OSWorld 2.0 eval separates the two tiers by 12.4 points (62.6% vs 50.2%) (ARC Prize, 2026-07-12; OpenAI GPT-5.6 announcement, 2026-07-12).
• Token economics complicate the "half the price" story. In CodeRabbit's harness, Terra used about 55,594 average output tokens per task against Sol's 20,968 — roughly 2.6x more — while solving fewer tasks; per-token price alone does not predict per-task cost (CodeRabbit blog, 2026-07-12).

Agentic behavior notes

• Programmatic Tool Calling. Terra can emit JavaScript that executes in an isolated V8 runtime with no network access, at the same capability level as Sol and Luna — relevant for agent harnesses that want sandboxed code execution without granting live network access (OpenAI API model docs; MarkTechPost, 2026-07-09).
• Higher tendency to exceed user intent. OpenAI's system card documents that all three GPT-5.6 tiers, Terra included, show an increased tendency versus GPT-5.5 to take agentic-coding actions the user did not request, though OpenAI describes absolute rates as low (Deployment Safety Hub system card, 2026-07-09).
• Independent read is more skeptical on capability, not on the behavior pattern. Zvi Mowshowitz's review of the same system card reports that Terra "tracks closer to GPT-5.5" than to Sol on most benchmarks, including cybersecurity- and biological-task measures, and that METR's external evaluation found only modest cyber-offense gains for Sol over GPT-5.5 — but he judges Terra to inherit the same architectural evaluation-awareness and situational-awareness concerns documented for Sol, just at lower absolute capability (Zvi Mowshowitz, 2026-06-28).
• Preparedness Framework rating. Terra is independently rated "High" (below "Critical") on both Cybersecurity and Biological/Chemical risk — OpenAI states this is the first time it has rated a smaller/faster tier this high (Deployment Safety Hub system card, 2026-07-09).

Limitations

• Vendor-vs-independent divergence on the flagship coding benchmark. CodeRabbit's own harness put Terra well below OpenAI's self-reported SWE-bench Pro number (40.7% pass rate vs. 63.4%) while consuming more output tokens per task than Sol, the opposite of the efficiency story the tier is marketed on (CodeRabbit blog, 2026-07-12).
• Alias trap. The bare gpt-5.6 model ID silently routes to Sol, not Terra — an easy way to overpay when trying to pin the cheaper tier (OpenAI API model docs).
• No fine-tuning support as of 2026-07-12 (OpenAI API model docs).
• Unverified comparative marketing claim. OpenAI states Terra and Luna "outperform Claude Fable 5 at around one-sixteenth the cost" on a vendor-defined 55-field professional-workflow benchmark; this research pass found no independent reproduction of that specific claim — treat it as an OpenAI claim, not an adjudicated fact (Willison, quoting OpenAI, 2026-07-12).
• Low-confidence tooling report. A single OpenAI community-forum post describes Terra in Codex CLI repeating the same status lines during a task at max reasoning effort, with no official OpenAI response on record — one anecdotal, single-source data point, not a confirmed defect (OpenAI community forum, 2026-07-09).
• Not selectable in plain ChatGPT. Terra is absent from the standard ChatGPT model picker on every plan; reaching it means going through ChatGPT Work (Plus/Pro/Business/Enterprise) or Codex, where it is the only tier available to Free/Go plans and one of three choices on Plus/Pro/Business/Enterprise (OpenAI Help Center).

Related

• gpt-5-6-sol — flagship tier of the same family; the comparison point for every row in the benchmark table above.
• gpt-5-6-luna — the cheaper, faster sibling tier.
• gpt-5-5 — the generation Terra is benchmarked against and is meant to match at lower cost.
• gpt-5-3-codex-spark — an earlier OpenAI coding-focused model, useful lineage context for Terra's agentic-coding positioning.
• claude-opus-4-8, claude-sonnet-5 — cross-vendor peers that appear alongside Terra in third-party benchmark and pricing comparisons.
• Context window, Context compaction — Terra's 1.05M-token window puts context-management technique choices in play for long agent sessions.